Anthropic shipped Claude Fable 5.1 on September 1. I ran it last night on the Max plan. About an hour in, the five-hour usage window was gone.
The output is good. I have not loved the usage limits. I am still planning in Claude and handing the build to Codex, because a full night of Fable doing both does not fit in the window.
What Anthropic actually shipped
Fable 5.1 and Mythos 5.1 are the same underlying model with different safeguards. Fable is the one in the model picker. Mythos stays inside Anthropic's trusted-access programs for cyber defense and life sciences. If you are building websites and automations, you are talking about Fable.
The API sticker did not move. Input is still $10 per million tokens. Output is still $50. Cache reads dropped 75 percent, to $0.25 per million. Anthropic's estimate from the launch post: typical work billed by token about 25 percent cheaper than Fable 5, and up to about 45 percent on long agentic runs where cache reads dominate the bill.
Artificial Analysis, which helped with pre-release testing, did not buy the savings story at max effort. The Decoder reported their numbers the same day: Fable 5.1 used about 1.7 times as many output tokens as Fable 5 on that setting, so the per-task cost on their Intelligence Index went up about 20 percent, to $3.76 versus $2.34 for Opus 5. Cheaper cache reads are real. A cheaper session is not automatic if the model thinks longer.
Claude Code needs version 2.1.255 or later to pick Fable 5.1. Default effort is High in Claude Code and Medium on claude.ai. High is the expensive setting. That matters more than the model name on a Max plan.
The benches that moved
Anthropic's own table, with production safeguards on:
- Terminal-Bench 4.0 (agentic coding): Fable 5.1 at 55.8 percent, Mythos 5.1 at 60.9, Fable 5 at 42.0, Opus 5 at 52.3, GPT-5.6 Sol at 37.3.
- Terminal-Bench-Science 0.1: 52.6 percent against 24.7 for Fable 5, 29.0 for Opus 5, and 22.4 for Sol. Biggest jump on the sheet.
- CursorBench 3.2.0: 73.4 percent, a few points over Fable 5, Opus 5, and Sol.
- AutomationBench: 31.4 percent, up from 17.1 on Fable 5.
Independent benches are less of a victory lap. CodeRabbit's September 1 review found almost the same known issues as Fable 5 (61.0 percent recall versus 61.9) with better precision (37.3 versus 32.8) and a lot less noise: 166 final comments instead of 253, and 79 nitpick-style comments instead of 265. The cost was time. Average review task took 18 minutes 38 seconds, against 12 minutes 32 seconds on Fable 5. Turning reasoning up from Low to High made it slower and found fewer issues.
Snorkel ran frontier coding tasks against Opus 5. On the matched set, both models solved 18 tasks, Opus alone solved 5, Fable alone solved 2. When Fable finished, the median successful run used 58 percent fewer output tokens and finished 36 percent faster. When it did not, the miss was often a last-mile thing: a lost terminal, a check it claimed to run and did not, or a build-and-dependency task. That category sat at 18 percent for Fable against 67 percent for Opus.
Amp switched its ultra mode to Fable 5.1 the same day. Their note is the one I trusted most on writing: they had not liked AI prose, and this version changed that. They also said it takes direction mid-task without dropping the original job, and that ultra threads ran about 35 percent cheaper because cache reads are most of an Amp thread. One internal example: typing latency in Safari on their iOS app went from 85 ms to 8 ms after a two-hour unattended pass.
Two days of reaction
The early-access quotes on Anthropic's post are the usual launch chorus. Jane Street said it stayed readable on long multi-step work. Cognition said they were moving Devin traffic off Opus 5 on launch day. Canva's head of AI preferred the writing in a blind test against Fable 5. Treat those as launch quotes.
The public timeline is louder about the meter. Kaushik Sridasyam wrote that running Fable 5.1 in subagents burned the five-hour limit and 27 percent of weekly usage in about 30 minutes. Another Max user said 15 minutes was enough to sit and wait for a reset. A $200-plan user called the consumption rate unusable for serious development work, even against Fable 5 on the same task. Those are posts, not lab results. They match what I saw.
That complaint is not new. Fable 5 already had a reputation for eating the window. PCMag's Sascha Segan wrote in July that he hit the five-hour cap constantly with Fable 5, often across a few prompts. Fable 5.1 did not fix that feeling. High effort in Claude Code made it easier to notice.
Why an hour ate a five-hour window
Anthropic's help article, updated September 2, is blunt. On Max, Fable 5 and Fable 5.1 are included. You can spend up to 50 percent of your weekly usage limits on them at no extra charge. They draw from the same weekly pool as every other Claude model, and they use that pool faster. The 50 percent is a ceiling, not a bonus. After you hit it, you keep going on usage credits or you switch models.
That is the weekly cap. Separately, every paid Claude plan still has a rolling five-hour session window. That is the one I hit. One hour of wall-clock time. A fan-out of parallel work. Session dead. Uncommitted work gone with it.
I already wrote about what happens when Anthropic gates the expensive model behind extra usage, and about compute partnerships that were supposed to give Claude Code more runway. Fable 5.1 is the same argument with a better model attached. The product is the intelligence. The constraint is still the window.
If you are going to spend Fable minutes, spend them on a spec or a review. Do not spend them on six agents writing files at once. That shape is how you turn a five-hour window into an hour, and an hour into a reset screen.
Plan in Claude, build in Codex
The output is why I still want Fable in the loop. It holds a messy repo in its head and writes a spec I can actually hand to someone. I have been doing that on purpose: plan and review in Claude, then give the build to Codex.
Codex spends its own budget. It keeps its own context. When the build seat dies, the planning seat is still alive. That sounds obvious written down. It was not obvious the night I let Fable try to do both.
We already looked at arena scores that had Codex slightly ahead of Claude Code on short coding tasks. Arena tests measure completions. They do not measure "write the spec, keep the session, and still have a working window at 10 p.m." For agency work, that second test is the one that decides whether a model is usable as a daily driver.
So the split, for now:
- Claude / Fable 5.1 for the plan, the review, and the hard diagnosis. Short sessions. One job at a time. No parallel fan-out unless I am ready to lose the window.
- Codex for the PR. Its own meter. A short report back. Claude's context stays for review, not for holding diffs.
Opus 5 is still the cheaper Claude for a lot of that planning if Fable is going to torch the cap. I will use Fable when the problem is actually hard. I will not use it as the default builder just because the picker put it on top.
I like Fable 5.1. I am not making it the all-day model on Max. Point it at the spec. Let Codex write the files. That is the version of this model I can actually finish a night of work with.
